Operator learning methods have found widespread applications in the field of scientific computing. However, existing methods often require a significant amount of high-fidelity data for convergent training, resulting in substantial data acquisition costs. In this paper, we propose a novel approach, named Multi-Fidelity Operator Learning (MFOL), for enhancing the performance of operator learning in scenarios where high-fidelity data is limited. In response to the scarcity of high-fidelity data, MFOL approximates high-fidelity solutions by leveraging multi-fidelity data. Specifically, MFOL incorporates a low-fidelity module designed to learn solutions from low-fidelity dataset. To bridge the gap between low and high fidelity, MFOL introduces a multi-fidelity module for capturing correlations across different fidelity data. In this context, we introduce the concept of multi-fidelity similarity to quantify these correlations. Additionally, we propose a data enhancement strategy based on operator learning to enhance the effective utilization of high-fidelity data within multi-fidelity learning frameworks. Experimental results on four equations demonstrate that our approach achieves a maximum of 33.7% improvement in prediction accuracy.